Time-Series Foundation Models for Forecasting Soil Moisture Levels in Smart Agriculture

TL;DR

This work investigates forecasting soil water potential $ψ_{soil}$ in smart agriculture using TimeGPT, a time-series foundation model. By evaluating zero-shot and fine-tuned variants (with and without exogenous inputs) against baselines like TFT, the study demonstrates that TimeGPT can achieve competitive accuracy, especially when fine-tuned on the target series history, while requiring less data and computation than end-to-end deep models. The results highlight the practical potential of foundation models for data-scarce agricultural settings and reveal that incorporating exogenous variables may not always improve performance, likely due to pre-training data distribution. Overall, the approach offers a data-efficient, scalable tool for irrigation decision support, aligning with sustainable development goals and real-world farming needs.

Abstract

The recent surge in foundation models for natural language processing and computer vision has fueled innovation across various domains. Inspired by this progress, we explore the potential of foundation models for time-series forecasting in smart agriculture, a field often plagued by limited data availability. Specifically, this work presents a novel application of $\texttt{TimeGPT}$, a state-of-the-art (SOTA) time-series foundation model, to predict soil water potential ($ψ_\mathrm{soil}$), a key indicator of field water status that is typically used for irrigation advice. Traditionally, this task relies on a wide array of input variables. We explore $ψ_\mathrm{soil}$'s ability to forecast $ψ_\mathrm{soil}$ in: ($i$) a zero-shot setting, ($ii$) a fine-tuned setting relying solely on historic $ψ_\mathrm{soil}$ measurements, and ($iii$) a fine-tuned setting where we also add exogenous variables to the model. We compare $\texttt{TimeGPT}$'s performance to established SOTA baseline models for forecasting $ψ_\mathrm{soil}$. Our results demonstrate that $\texttt{TimeGPT}$ achieves competitive forecasting accuracy using only historical $ψ_\mathrm{soil}$ data, highlighting its remarkable potential for agricultural applications. This research paves the way for foundation time-series models for sustainable development in agriculture by enabling forecasting tasks that were traditionally reliant on extensive data collection and domain expertise.

Figures (6)

• Figure 1: Top: conventional approaches typically require large amounts of data to train a well-performing forecasting model, which is not always feasible in agriculture due to high costs involved in sensor setup, maintenance, ... Bottom: foundation models require no data at all for inference (in a zero-shot setting) or require far less data for fine-tuning compared to conventional approaches.
• Figure 2: Overview of the $\texttt{TimeGPT}$ architecture based on garza2023timegpt. Note that $\mathbf{X}$ can be univariate (containing only the history of the target), or multivariate (containing exogenous variables along with the history of the target).
• Figure 3: Overview of the datasplits into training, validation, and test set. Final performance is reported on the test set.
• Figure 4: Forecasts from the best performing models and the naive for three randomly sampled sensors across different orchards. The error bands for $\texttt{TimeGPT}$ and the TFT are omitted for clarity.
• Figure 5: Correlation plots across the forecasting horizon $h$ representing the correlation between $\psi_\mathrm{soil}$ at time $t$ and $\psi_\mathrm{soil}$ at $t+1$ to $t+5$. Note how the error becomes larger as time moves on.